arch/ia64/lib/copy_page.S

   1 /*
   2  *
   3  * Optimized version of the standard copy_page() function
   4  *
   5  * Inputs:
   6  *      in0:    address of target page
   7  *      in1:    address of source page
   8  * Output:
   9  *      no return value
  10  *
  11  * Copyright (C) 1999, 2001 Hewlett-Packard Co
  12  *      Stephane Eranian <eranian@hpl.hp.com>
  13  *      David Mosberger <davidm@hpl.hp.com>
  14  *
  15  * 4/06/01 davidm       Tuned to make it perform well both for cached and uncached copies.
  16  */
  17 #include <asm/asmmacro.h>
  18 #include <asm/page.h>
  19
  20 #define PIPE_DEPTH      3
  21 #define EPI             p[PIPE_DEPTH-1]
  22
  23 #define lcount          r16
  24 #define saved_pr        r17
  25 #define saved_lc        r18
  26 #define saved_pfs       r19
  27 #define src1            r20
  28 #define src2            r21
  29 #define tgt1            r22
  30 #define tgt2            r23
  31 #define srcf            r24
  32 #define tgtf            r25
  33 #define tgt_last        r26
  34
  35 #define Nrot            ((8*PIPE_DEPTH+7)&~7)
  36
  37 GLOBAL_ENTRY(copy_page)
  38         .prologue
  39         .save ar.pfs, saved_pfs
  40         alloc saved_pfs=ar.pfs,3,Nrot-3,0,Nrot
  41
  42         .rotr t1[PIPE_DEPTH], t2[PIPE_DEPTH], t3[PIPE_DEPTH], t4[PIPE_DEPTH], \
  43               t5[PIPE_DEPTH], t6[PIPE_DEPTH], t7[PIPE_DEPTH], t8[PIPE_DEPTH]
  44         .rotp p[PIPE_DEPTH]
  45
  46         .save ar.lc, saved_lc
  47         mov saved_lc=ar.lc
  48         mov ar.ec=PIPE_DEPTH
  49
  50         mov lcount=PAGE_SIZE/64-1
  51         .save pr, saved_pr
  52         mov saved_pr=pr
  53         mov pr.rot=1<<16
  54
  55         .body
  56
  57         mov src1=in1
  58         adds src2=8,in1
  59         mov tgt_last = PAGE_SIZE
  60         ;;
  61         adds tgt2=8,in0
  62         add srcf=512,in1
  63         mov ar.lc=lcount
  64         mov tgt1=in0
  65         add tgtf=512,in0
  66         add tgt_last = tgt_last, in0
  67         ;;
  68 1:
  69 (p[0])  ld8 t1[0]=[src1],16
  70 (EPI)   st8 [tgt1]=t1[PIPE_DEPTH-1],16
  71 (p[0])  ld8 t2[0]=[src2],16
  72 (EPI)   st8 [tgt2]=t2[PIPE_DEPTH-1],16
  73         cmp.ltu p6,p0 = tgtf, tgt_last
  74         ;;
  75 (p[0])  ld8 t3[0]=[src1],16
  76 (EPI)   st8 [tgt1]=t3[PIPE_DEPTH-1],16
  77 (p[0])  ld8 t4[0]=[src2],16
  78 (EPI)   st8 [tgt2]=t4[PIPE_DEPTH-1],16
  79         ;;
  80 (p[0])  ld8 t5[0]=[src1],16
  81 (EPI)   st8 [tgt1]=t5[PIPE_DEPTH-1],16
  82 (p[0])  ld8 t6[0]=[src2],16
  83 (EPI)   st8 [tgt2]=t6[PIPE_DEPTH-1],16
  84         ;;
  85 (p[0])  ld8 t7[0]=[src1],16
  86 (EPI)   st8 [tgt1]=t7[PIPE_DEPTH-1],16
  87 (p[0])  ld8 t8[0]=[src2],16
  88 (EPI)   st8 [tgt2]=t8[PIPE_DEPTH-1],16
  89
  90 (p6)    lfetch [srcf], 64
  91 (p6)    lfetch [tgtf], 64
  92         br.ctop.sptk.few 1b
  93         ;;
  94         mov pr=saved_pr,0xffffffffffff0000      // restore predicates
  95         mov ar.pfs=saved_pfs
  96         mov ar.lc=saved_lc
  97         br.ret.sptk.many rp
  98 END(copy_page)